Graph Deep Learning for Time Series Processing
Forecasting, Reconstruction and Analysis
Successful applications of deep learning in time series processing often involve training a single neural network on a collection of (related) time series. Pairwise relationships among time series can be modeled by considering a (possibly dynamic) graph spanning the collection. In this context, graph-based methods take the standard deep learning approach to time series processing a step forward. The recent theoretical and practical developments in graph machine learning make adopting such an approach particularly appealing and timely. The twofold objective of this tutorial is to:
- offer a comprehensive overview of the field, with a focus on forecasting applications;
- provide tools and guidelines to design and evaluate graph-based models for time series.
Material
Download the slides used in our tutorial.
This tutorial will be delivered at the Learning on Graphs (LoG) Conference, held online from the 26th to the 29th of November 2024, and at the Italy Meetup, in Siena, Italy, from the 4th to the 6th of December 2024.
The tutorial will take place on Thursday, 28th of November, 17:00-20:00 (London, UTC), and Friday, 6th of December, 10:30-12:00 (Rome, UTC+1).
Program
Part 1 Graph deep learning for time series processing.
- Correlated time series
Definition of the problem settings. Time series forecasting. - Graph deep learning for time series forecasting
Graph-based framework for representing correlated time series. Graph-based models for time series forecasting. Similarities to and differences from related settings in time series analysis and temporal graph learning. - Spatiotemporal graph neural networks (STGNNs)
Core components of a STGNN model. Recipes and strategies for building STGNNs. The time-then-space and time-and-space paradigms. Overview of architectures from the literature. - Global and local models
Parameter sharing in time series models. Review of the global and local modeling paradigms with their strengths. Practical implications in graph-based time series processing. Hybrid global-local STGNN architectures. Transfer learning.
Demo Coding Spatiotemporal GNNs
Overview of open-source Pytorch libraries for graph-based time series processing. Torch Spatiotemporal demo.
Part 2 Challenges and tools.
- Latent graph learning
Methods to apply the framework when no pre-defined graph is available. Learning graph structures from collections of time series. - Scalability
Methods to enable scalability to large sensor networks. - Dealing with missing data
The problem of missing data. Overview of methods for graph-based multivariate time series imputation and kriging. - Model quality assessment
Statistical tools to test the optimality of graph-based predictors. Identification of time-space regions where predictions can be improved.
Organizers
This tutorial is organized by the GMLG Research Group within the Swiss AI Lab IDSIA and Università della Svizzera italiana.
